How MRT Delays Propagate and Recover Using Mathematics: How One Lost Minute Travels Through a Railway

An MRT delay does not stay where it begins.

A door is held for fifteen extra seconds. A train approaches a station more slowly than planned. Passenger exchange takes longer than expected. A temporary operational constraint changes one movement.

At first, the disturbance may look tiny.

But a metro is a sequence of trains, not a collection of independent vehicles.

If Train B is delayed, the gap in front of B becomes larger while the gap behind B becomes smaller. More passengers can accumulate before B. Fewer may accumulate before Train C. B can become more crowded and take longer at the next station, while C can catch up.

The delay has changed shape.

It is no longer merely:

Train B is 15 seconds late

It has become:

headway distortion
+ passenger redistribution
+ dwell-time change
+ following-train interaction
+ network recovery problem

A railway delay is not simply lost time. It is a disturbance travelling through a moving system.

This article continues the eduKateSG Singapore MRT mathematics series. Begin with How MRT Timing Works Using Mathematics, then see our articles on MRT braking, MRT headway, and station dwell time.

The RFE — Why Does Delay-Recovery Mathematics Exist?

No real railway can guarantee that every train will reproduce the planned timetable to the second forever.

Reality varies.

• Passenger loads vary.
• Dwell times vary.
• Running times vary.
• Equipment and operating conditions vary.
• Occasional incidents and faults occur.

So a railway needs more than a plan.

It needs the ability to return towards the plan — or, when the original timetable is no longer immediately achievable, towards a stable and useful operating pattern.

The Reason for Existence is therefore:

detect deviation
→ prevent unsafe interaction
→ prevent local delay becoming network instability
→ restore regular useful service
→ reduce passenger cost
→ return the railway towards a stable state

This is also a live Singapore concern. In January and February 2026, LTA, SMRT and SBS Transit announced enhanced rail service-recovery and commuter-management measures following the Rail Reliability Taskforce’s recommendations, including stronger contingency planning and procedures intended to support quicker and more effective service resumption after faults.

Those public measures concern real disruption recovery. In this article, we focus on the underlying mathematics of how disturbances propagate and how railway systems can conceptually recover, without attempting to reproduce any operator’s proprietary control rules.

Prompt 1 — What Is a Delay Mathematically?

Suppose train i is scheduled to depart station k at time:

Dᵢ,ₖˢ

but actually departs at:

Dᵢ,ₖᵃ

Then departure delay is:

δᵢ,ₖ = Dᵢ,ₖᵃ − Dᵢ,ₖˢ

If:

δ = 0

the train is exactly on the reference timetable at that point.

If:

δ = +20 s

the train is twenty seconds late.

This looks simple.

But high-frequency metro operation often cares just as much about relative spacing as absolute timetable error.

Define headway between Train i and the previous train:

Hᵢ,ₖ = Dᵢ,ₖ − Dᵢ₋₁,ₖ

If desired headway is H*:

eᵢ,ₖ = Hᵢ,ₖ − H*

is headway error.

A train can therefore have two different kinds of “wrongness”:

• it can be late relative to the timetable,
• and it can be badly spaced relative to neighbouring trains.

These are not always the same problem.

Example: late but regular

Suppose every train on a section is delayed by exactly 60 seconds:

A +60
B +60
C +60
D +60

The timetable is shifted.

But headways may remain perfectly regular.

Example: near timetable but irregular

Now suppose delays are:

A +0
B +30
C −20
D +10

The average delay may not look dramatic.

But spacing can be badly distorted.

For a high-frequency metro, regularity can sometimes matter more to passengers than whether every train matches an old clock time exactly.

Prompt 2 — What Is the Difference Between Primary and Secondary Delay?

Railway-delay research often distinguishes between an initial disturbance and delays caused by its interaction with the rest of the timetable.

We can call the first disturbance a primary delay.

Example:

Train B loses 30 seconds at Station 4

The consequences created afterwards are secondary delays.

For example:

Train C slows because B is closer ahead
Train B takes longer at Station 5 because more passengers accumulated
Train D approaches a junction at a less favourable time

These effects did not cause the original problem.

They are consequences of the first disturbance interacting with the system.

A conceptual decomposition is:

Total delay
=
primary delay
+ secondary delay

This distinction matters because recovery is not only about removing the original fault.

Even after the original cause disappears, the railway may still be carrying the disturbance downstream.

The cause can be gone while the consequences are still travelling.

This is why modern rail service recovery is a system problem, not merely a repair problem.

Prompt 3 — How Does One Late Train Change Two Headways?

Imagine three trains:

A → 120 s → B → 120 s → C

Desired headway is 120 seconds.

Now Train B loses 20 seconds.

Its departure becomes later.

The pattern becomes approximately:

A → 140 s → B → 100 s → C

One 20-second delay has created:

• a 20-second excess gap in front of B,
• and a 20-second shortage of gap behind B.

The mean of the two gaps remains:

(140 + 100)/2 = 120 s

If we looked only at average headway, we could mistakenly conclude that nothing changed.

But the variance changed.

Let mean headway be H̄. A simple variance measure is:

Var(H) = (1/n) Σ(Hᵢ − H̄)²

Before the delay:

Var(H) = 0

Afterwards, it is positive.

The railway has moved from regular to irregular even though its average separation remained 120 seconds.

This is why our headway article emphasised that passengers pay for variance.

Prompt 4 — Why Can the Late Train Become Even Later?

Now passengers enter the equation.

Let λ be passenger arrival rate at a platform.

Expected passenger accumulation during headway H is approximately:

N ≈ λH

Suppose:

λ = 2 passengers/s

At the original 120-second gap:

N = 240 passengers

After Train B’s delay creates a 140-second gap:

N = 280 passengers

Forty additional passenger arrivals are now associated with B’s larger gap in this simple model.

More boarding demand may produce a longer dwell.

Longer dwell may make B later still.

Meanwhile the 100-second gap before C produces:

N = 200 passengers

C may therefore experience less boarding demand and a shorter dwell.

C catches B.

The feedback loop becomes:

B delayed
→ gap before B grows
→ more passengers accumulate
→ B dwell increases
→ B delayed more

C gap shrinks
→ fewer passengers accumulate
→ C dwell decreases
→ C catches B

This is the mathematics behind train bunching.

Delay can feed itself because passenger demand responds to the delay.

Research on automatic train regulation specifically notes this instability in high-frequency metros: schedule deviations can amplify because the delayed train encounters accumulated passengers, which then affects dwell and subsequent movement.

Prompt 5 — How Does Delay Travel from Station to Station?

Let δ_k be the delay of a train after station k.

A useful educational recurrence is:

δₖ₊₁
=
δₖ
+ ΔTdwell,k
+ ΔTrun,k
− Rₖ

where:

• ΔT_dwell,k = extra dwell delay at station k,
• ΔT_run,k = extra running-time delay on the next section,
• R_k = time recovered from available operating margin or regulation.

If:

ΔTdwell + ΔTrun > R

delay grows.

If:

ΔTdwell + ΔTrun < R

delay shrinks.

If they are equal, delay persists unchanged.

This recurrence turns the railway into a discrete dynamical system.

One state feeds the next:

δ₁ → δ₂ → δ₃ → δ₄ → ...

Now add passenger state P_k and headway state H_k:

δₖ₊₁ = f(δₖ, Hₖ, Pₖ, uₖ, wₖ)

where:

• u_k represents control action,
• w_k represents random disturbance or uncertainty.

Now we have the essential structure of a stochastic control system.

Delay as a wave

We can also think spatially.

A disturbance begins at one station and time:

(station 4, 08:12)

Later it may appear at:

station 5, 08:14
station 6, 08:16
station 7, 08:18

and may also affect following trains.

Metro research therefore studies disturbance propagation in both:

space
and
time

A 2025 study in Control Engineering Practice, for example, models the spatial-temporal propagation and recovery of supply-demand imbalance in metro systems, tracking where and when additional passenger waiting cost appears after a disturbance.

The delay is therefore not only a number.

It has a location, a direction and a changing footprint.

Prompt 6 — What Is Recovery Margin?

A timetable does not necessarily schedule every interstation run at the absolute physical minimum time.

Suppose a train can safely and comfortably complete a section under favourable conditions in:

Tmin = 72 s

but the timetable allocates:

Tschedule = 78 s

The difference:

r = Tschedule − Tmin
  = 6 s

is a simplified example of running-time supplement or recovery margin.

If a train is four seconds late and conditions permit it to use four seconds of that margin:

δbefore = 4 s
R = 4 s
δafter ≈ 0

The delay disappears without requiring the train to violate speed, braking or comfort limits.

This is why a timetable that appears slightly slower than the theoretical minimum can actually produce a faster passenger experience over a full day.

It is more robust.

Slack is not always wasted time. Properly placed slack is stored recovery capability.

Railway delay-propagation research explicitly studies both running-time supplements and headway buffers because these margins influence how much of an initial disturbance becomes secondary delay.

Prompt 7 — How Can Mathematics Pull Trains Back into Rhythm?

Once trains are irregular, there are conceptually several levers a metro operator may use, depending on the system, safety rules and operating situation.

Academic metro-control research commonly studies tools such as:

• speed regulation between stations,
• holding a train briefly at a station,
• adjusting departure order or timetable decisions,
• short-turn or other rescheduling strategies during larger disruptions,
• and joint control of passenger flow and train regulation.

This does not mean every strategy is used in the same way on Singapore’s MRT. The purpose here is to understand the mathematics shared by high-frequency metro systems.

Speed regulation

Suppose Train B is late but still has available running-time margin.

Its interstation trajectory might be adjusted within all permitted constraints so that it uses less coasting or a different speed profile and recovers some time.

Conceptually:

Trun,new = Trun,scheduled − r

where r is recoverable time.

Holding

Holding sounds counterintuitive.

Why deliberately make a train wait when the railway is already late?

Because the goal may be regularity rather than the speed of one individual train.

Suppose:

A → 150 s → B → 90 s → C

If C catches B too aggressively, the pair can bunch.

A small control hold on C can prevent the short gap becoming even shorter.

That may slightly delay C while reducing a larger system cost.

This is optimisation:

accept small local cost
→ avoid larger network cost

Passenger-flow control

When platforms are overloaded, the railway problem is no longer only train movement.

Passenger inflow itself can become part of the state being managed.

Research on high-frequency metro systems develops joint models that regulate train movement and passenger flow together because the two systems are dynamically coupled.

The key insight is:

trains influence passengers
and
passengers influence trains

Recovery therefore requires both sides of the loop to be understood.

Prompt 8 — What Does an Optimal Recovery Problem Look Like?

We can now write a conceptual control objective.

Let:

• δ_i,k = timetable delay,
• e_i,k = headway error,
• W_k = passenger waiting cost,
• U_i,k = control effort or deviation from normal operation.

A simplified objective might be:

minimise J
=
Σ [
w1 δᵢ,ₖ²
+ w2 eᵢ,ₖ²
+ w3 Wₖ
+ w4 Uᵢ,ₖ²
]

subject to constraints:

safe separation maintained
speed limits respected
braking envelopes respected
dwell ≥ safe passenger requirement
available rolling stock respected
junction conflicts avoided
terminal constraints respected

The weights w₁, w₂, w₃ and w₄ represent the relative importance assigned to different operating costs.

This reveals a critical idea.

The goal is not necessarily:

make every delayed train individually as fast as possible

because that can worsen bunching or create conflicts elsewhere.

The goal is closer to:

restore the best feasible system state

for the railway and the passengers together.

Recovery is a network optimisation problem disguised as a late train.

A Worked Delay-Propagation Example

Consider four hypothetical trains A, B, C and D operating at a desired 120-second headway.

Initial pattern:

A →120→ B →120→ C →120→ D

At Station 1, Train B experiences a 30-second excess dwell.

Immediately after departure:

A →150→ B →90→ C →120→ D

Now suppose passenger arrival rate is 2 passengers per second.

Passenger batches associated with the first two gaps are approximately:

before B: 2 × 150 = 300
before C: 2 × 90  = 180

Compared with the original 240 passengers per 120-second gap:

B sees +60 passenger arrivals
C sees −60 passenger arrivals

Suppose the larger passenger exchange adds another 8 seconds to B’s dwell at Station 2.

But B has 6 seconds of recoverable running-time margin.

Net new delay:

+8 − 6 = +2 s

B’s delay becomes:

32 s

The disturbance is still growing, but slowly.

Now suppose the next station has lighter demand. B gains no extra dwell and recovers 6 seconds.

32 − 6 = 26 s

The system has entered recovery.

Over successive stations, we could obtain:

δ₁ = 30
δ₂ = 32
δ₃ = 26
δ₄ = 18
δ₅ = 9
δ₆ = 2
δ₇ = 0

The sequence shows a disturbance that first worsens, then decays.

But now imagine every station is crowded enough to add new dwell faster than the train can recover.

The sequence could instead be:

30, 36, 44, 55, 69, ...

The railway has moved into an unstable regime.

The same original 30-second delay can therefore have very different futures depending on passenger demand, margin and control.

Stable and Unstable Delay Mathematics

We can create an extremely simplified linear model:

δₖ₊₁ = aδₖ

If:

|a| < 1

delay decays over time.

For example:

a = 0.8
30 → 24 → 19.2 → 15.4 → ...

If:

a = 1

delay persists.

If:

a > 1

delay grows.

For example:

a = 1.2
30 → 36 → 43.2 → 51.8 → ...

A real railway is not described by one constant a, but this tiny model teaches the concept of stability.

The recovery question becomes:

does the system naturally reduce deviations,
or do deviations amplify?

Control is the deliberate attempt to move the system towards the first behaviour.

The Difference Between Timetable Recovery and Service Recovery

These terms should not be confused.

Timetable or regularity recovery

This can describe the smaller-scale mathematical problem of returning trains towards planned times or planned headways after ordinary disturbances.

Service recovery after a disruption

This is the larger operational problem after a fault or major interruption. It may involve restoring train service, managing commuters, deploying contingency plans and reconnecting the line to normal operation.

Singapore’s Rail Reliability Taskforce recommendations and the January–February 2026 LTA announcements refer to this broader service-recovery problem.

Mathematically the two are related, but they operate at different scales.

small disturbance
→ regulate timing/headway

major disruption
→ restore usable service
→ manage passenger network
→ then restore normal timetable/regularity

One should not be mistaken for the other.

Why Passenger Cost Changes the Best Recovery Strategy

Suppose two recovery strategies produce the same total train delay.

Are they equally good?

Not necessarily.

Strategy A may leave one lightly loaded train badly delayed.

Strategy B may produce smaller delays across several extremely crowded trains.

Train-delay minutes alone can hide the passenger effect.

So we might define passenger-weighted delay:

PD = Σ Pᵢ δᵢ

where P_i is the number of affected passengers associated with train i.

Now a one-minute delay on a heavily loaded train counts differently from a one-minute delay affecting few passengers.

Research on real-time metro rescheduling increasingly includes passenger satisfaction, passenger flows and waiting cost directly in the optimisation objective.

The railway carries trains physically, but the purpose of recovery is to recover people’s journeys.

The Recovery Horizon Problem

How far ahead should a recovery algorithm look?

If it looks only at the next station, it may make a locally good choice that causes trouble five stations later.

If it tries to optimise the entire day in full detail every second, the computational problem may become unnecessarily large.

So dynamic transport optimisation often uses a prediction horizon.

At time t:

observe current state
→ predict next N minutes
→ choose control action
→ apply first action
→ observe again
→ solve again

This is closely related to model predictive control.

The railway never assumes its prediction will remain perfect forever.

It repeatedly receives new reality.

predict
→ act
→ observe
→ update
→ predict again

This is the same World Return logic we saw in the main MRT timing article, expressed as control mathematics.

Why Recovery Can Temporarily Look Worse Before It Looks Better

A recovery decision can impose a small new delay in order to eliminate a larger instability.

Suppose Train C is only 80 seconds behind B when desired headway is 120 seconds.

If C is allowed to keep catching B, they may become a pair.

A temporary 15-second hold could increase the gap to 95 seconds.

Another regulation later might increase it further.

During the first control action, C becomes later.

But system regularity improves.

This is why optimisation must distinguish:

local performance
from
system performance

A mathematically intelligent railway may sometimes make one train slightly worse so that the railway becomes much better.

Recovery as a Phase Transition

There is a useful way to think about system states.

STATE 1 — stable normal operation
small disturbances decay

STATE 2 — stressed operation
small disturbances persist

STATE 3 — unstable operation
disturbances amplify and bunching grows

STATE 4 — disrupted operation
normal timetable no longer describes useful service

The job of recovery is not always immediately:

STATE 4 → STATE 1

Sometimes the practical path is:

STATE 4
→ establish safe limited service
→ stabilise passenger and train flow
→ rebuild regularity
→ return to normal operation

This layered view helps explain why “the fault has been fixed” does not necessarily mean “the railway is instantly normal”.

Trains and passengers may still be displaced from their normal pattern.

A Simple Recovery Metric

Suppose system disturbance magnitude is measured by:

E(t) = Σ eᵢ(t)²

where e_i is headway error for each train.

If:

E(t) → 0

the train sequence is returning towards uniform headway.

We can define recovery time:

Trecover
=
first time after disturbance
when E(t) remains below threshold ε

This makes recovery measurable.

We are no longer merely saying:

things look normal again

We can define what “normal enough” means mathematically.

Why Resilience Is Different from Reliability

Reliability asks:

How often does the system avoid failure or major disruption?

Resilience asks:

When something does go wrong,
how well and how quickly can the system return?

These are related but different properties.

A useful conceptual resilience metric might combine:

disturbance magnitude
× disturbance duration
× passenger impact

If service quality is Q(t), one can think of total disruption loss as an area:

Loss = ∫ [Qnormal − Q(t)] dt

A resilient system reduces that area by:

• limiting how far service quality falls,
• shortening how long the reduction lasts,
• and reducing passenger harm while recovery occurs.

This gives a much richer view than counting whether an incident happened.

RFE Return — What Does It Mean for an MRT to Recover?

It does not necessarily mean:

every train instantly returns to its original second-by-second timetable

A more meaningful definition can be:

safe train movement restored
+ passenger flow manageable
+ headway irregularity shrinking
+ excessive queues dissipating
+ operational margins rebuilding
+ timetable alignment returning

Recovery is therefore a trajectory, not a switch.

The railway leaves equilibrium.

The system responds.

It observes what happened.

It adjusts again.

And eventually the disturbance becomes smaller than the system’s ability to absorb it.

A recovered railway is not one that never left the plan. It is one that knows how to return after reality pushes it away.

Conclusion — One Lost Minute Has a Shape

We often speak about delay as though it were a quantity sitting on one train.

“The train is one minute late.”

But one minute can have a shape.

It can stretch the gap before one train.

It can compress the gap behind it.

It can attract more passengers.

It can lengthen the next dwell.

It can force another train to slow.

It can move from station to station.

It can become ten minutes of passenger waiting spread across hundreds of people.

Or, if the railway has sufficient margin and intelligent control, it can gradually disappear.

disturbance
→ propagation
→ observation
→ control
→ recovery

That sequence may be one of the most important mathematical loops in public transport.

The passenger sees a late train.

The statistician sees a deviation.

The control engineer sees an error state.

The operations researcher sees a constrained recovery problem.

The station sees an altered passenger queue.

The city experiences the combined result.

And so the real question is not:

Can an MRT ever be delayed?

No real system can make that promise.

The deeper question is:

When time is lost, how well can the railway stop that loss from multiplying?

That is where delay mathematics becomes resilience.

Key Mathematical Ideas

δᵢ,ₖ = Dᵢ,ₖᵃ − Dᵢ,ₖˢ
Timetable delay

Hᵢ,ₖ = Dᵢ,ₖ − Dᵢ₋₁,ₖ
Headway

eᵢ,ₖ = Hᵢ,ₖ − H*
Headway error

Var(H) = (1/n)Σ(Hᵢ − H̄)²
Headway irregularity

N ≈ λH
Passenger accumulation

δₖ₊₁ = δₖ + ΔTdwell,k + ΔTrun,k − Rₖ
Simplified delay propagation

r = Tschedule − Tmin
Simplified recovery margin

δₖ₊₁ = aδₖ
Elementary stability model

PD = ΣPᵢδᵢ
Passenger-weighted delay

J = Σ[w1δ² + w2e² + w3W + w4U²]
Illustrative recovery objective

E(t) = Σeᵢ(t)²
Illustrative network irregularity measure

Loss = ∫[Qnormal − Q(t)]dt
Conceptual resilience loss

Important: These equations are educational abstractions. Actual Singapore MRT traffic regulation, disruption management and safety procedures use operator-specific systems, validated engineering rules and operational protocols that are not reproduced here.

Continue the MRT Mathematics Series

• How MRT Timing Works Using Mathematics — the full system pillar.
• How MRT Braking Works Using Mathematics — stopping distance, braking curves and energy.
• How MRT Headway Works Using Mathematics — how train spacing becomes capacity.
• How MRT Station Dwell Time Works Using Mathematics — passenger queues and doorway bottlenecks.

The next mathematical leg can move from recovery to the whole timetable: how an MRT line is scheduled before service begins, how cycle time determines fleet size, and how optimisation fits trains, stations, terminals and recovery margins into one operating day.

Sources and Further Reading

• Land Transport Authority — Enhanced Rail Service Recovery and Commuter Management Measures, 24 January 2026
• Land Transport Authority, SMRT and SBS Transit — Rail Reliability Taskforce Recommendations, 13 February 2026
• Control Engineering Practice — Data-driven propagation and recovery of supply-demand imbalance in a metro system
• Transportation Research Part C — Simulation-optimisation framework for metro disturbance recovery
• Transportation Research Part C — Analytical control of disruptions using holding and speed control
• Transportation Research Part B — Automatic train regulation with transfer coordination
• Transportation Research Part C — Railway line delay propagation, running-time supplements and headway buffers
• Transportation Research Part B — Joint train regulation and passenger-flow control for high-frequency metro lines
• Transportation Research Part B — Real-time train scheduling during metro disruptions

Return to the complete MRT Mathematics hub: How MRT Works | It’s Mathematics.